|
This edition of TMAI Premium is free for Standard Edition subscribers.Â
Over the last few weeks, Premium editions have unpacked high-impact topics. 1. How to build a large-scale experimentation platform? 2. The five game-changing CEO-Level KPIs - for profit & customer delight.  3. Identifying the crucial why using the simple 3Q Survey (to complement the quantitative what). 4. How to conduct high-quality surveys and analysis? You can upgrade to TMAI Premium here. All Premium revenue is donated to charity. | | TMAI #378: Unlocking Machine Learning: Propensity Models - P1
| At some level, you get that every human your Owned, Earned, Paid Marketing initiatives engage with is going to worth different amounts in customer value. The challenge has always been that figuring out customer value is only possible AFTER you’ve already spent money reaching them (with hopefully a non-interruptive, hyper-relevant, piece of content). They have to come to your store/site/app. They have to stay. They have to actively absorb your value proposition. They have to add-to-cart. They have to check out. On average, approx. 2% of the humans do that. At this point, if you have an easy, awesome, connected CRM system, you can identify customer value for that 2%. (Not longer-term value, just initial value.) It would be magical if you knew this up front: Pooja’s customer value is $410. Sandeep, $30. Avinash, $0. Then, you could invest $40 on engaging Pooja, $0.5 on Sandeep, and $0 on me! Now imagine understanding that, and activating tactics, for the tens of thousands (/millions) of humans your Marketing engages every day. Can you see the incredible increase in effectiveness (profit) and efficiency (budget saved)? Yes. OMG! The quest to figure this out, up front? Propensity Modeling. The desire to do propensity modeling has existed for 2,000 years. The value is clear, the impact clearly huge. But, for 1,997 years, it has been painful, sketchy, and low-quality predictions. Especially for digital – yes, we have exponentially more date, it is also fragmented, it has holes, it is of variable anonymity, yada, yada, yada. Made it a challenge to read patterns. As with so many things these days, machine learning has changed the game. Today, with help from my peer Sal, the sexy story of the propensity modeling work we are doing at Croud. It is helping us deliver truly transformative client results. | Higher Order Context. You’ve heard me make a strong case against understanding customers using demographics (age, gender, etc.) and psychographics (education, income, geo, etc.). That approach prevents you from understanding that, as an example, I (a man) prefer women's sneakers because they are invariably designed better, in ravishing colors! Historically, you’ve not had a choice. Your quest to understand a customer was descriptive. Understanding the past behavior (if you had it well organized), and relying on clichés (demographics, psychographics). The explosion of machine learning algorithms, and their availability to us mere mortals at small companies, over the last three years, have resulted in a shift towards predictive analytics – allowing us to anticipate future consumers’ needs and behavior. Finally, we can have something smarter that can tease out patterns in fragmented data with holes and variable anonymity! | Propensity Modeling: Foundations.
Propensity models leverage machine learning to forecast the likelihood of a specific event to occur (propensity score) based on consumer features and historic patterns.
In our use cases, this also involves analyzing how the user interacts with a website or an app.
The model will produce a propensity score, which, in our use cases, indicates how likely the user is to make a purchase.
At Croud our journey started with propensity models trained to predict the (potential) value of (potential) customers.
This requires a more complex architecture, as the propensity model will need to be augmented to predict the likelihood of purchase AND the value of the purchase. In some instances, our models also help predict the more complex (but invaluable) frequency of purchases.
Simply put: Our models are trained to predict the value of the next purchase, or to estimate it over a longer time period (lifetime!).
[Note: Premium members, please review TMAI #369, #370, #371, LTV is Not Marketing’s Friend, for the critical worldview to bring to this quest. If you can’t find them,  email me.]
The immense benefit of the longer time frames, as the above Premium editions outline, is that there will be a dramatic impact on the way organizations strategize their relationship with a customer from acquisition to retention, using investments across departments - and, not just Marketing! | Propensity Modeling: The Croud Way. (B2B Use Case.)
Let me share a challenging use case.
For a B2B client, the site delivers a lead. Next steps occur in other departments (virtual Sales, in-person Sales). Days, or weeks, later some of these leads convert. All this data ends up in a different corporate system.
An immediate onsite conversion would be wonderful. In this case, our challenge was to predict the likelihood (and value) at the end of a longer sales cycle, offline. Then, use that likelihood and value to execute customer-value based bidding in PPC campaigns to target predicted high-value customers.
Because this is normally such a data engineering challenge, with lagging outcomes, your company will choose to do customer-value based bidding purely based on Leads Submitted. This is would suck less than not using that signal, but you will still end up targeting loads of Avinashs ($0), Sandeeps ($40) and maybe none of the Poojas ($410).
It would not be because Google or Facebook have bad algorithms. No. Their algorithms are pretty sweet. They are going to take your success signal, merge it with terabytes of their proprietary data, and make smart-real time decisions re bids and targeting. The problem: Your signal is super weak, in terms of value to your business.
This is the before version: |
|
We have delivered the advertising to a group of users, based on the best available signals related to demographics, psychographics, (or the best) using intent inferred through expressed behavior. [Refer to my See-Think-Do-Care framework.] We will only know after the fact, weeks after the lead is submitted, each of these were wroth different customer value. Hence, there is a chance we underbid for high-value customers (Pooja), and bid on loads of low-value ones (Avinash). Without divulging trade secrets (!), but with enough detail to help you get it, here’s the Croud way: |
|
We are leveraging website traffic data to analyze the behavior of each user on the website and measuring multiple factors related to engagement.
We are augmenting this data with first-party data available to us from the client’s CRM system (remember those days, weeks, later conversions?).
The combination of these two, using the algorithms above, results in a propensity score based on our prediction for both likelihood to convert and the conversion value (post the online, offline conversion process).
In a digital use case, this score gets passed back to, say, Google’s customer-value based bidding with benefiting from this superior intelligence will place smarter bids on behalf of our client (the second green column).
You remember how deeply in my soul I believe in trust, but verify when it comes to super advanced math (Premium members see TMAI #287). We measure that last column, and go back and check against our original predictions that powered the propensity score. This helps us fine-tune our models to ensure our algorithms become smarter over time.
The impact on business profits for our clients has been transformational. It’ll be the same for you/your clients.
Imagine the implications on a normal ecommerce experience.
Imagine killing all the initiatives around customer personas and customer segmentation (projects that can stretch years and millions of dollars of time and money), and letting a machine figure out who your low-value, medium-value, high-value customers are based on exponentially more data than you could ever imagine a human being able to ingest.
Imagine plugging additional signals around creatives to transform right message, in addition to right person and right time (which you see in the above picture).
The implications are super-exciting, as you internalize what's underpinning propensity modeling. | Next Week.
In next week's Premium edition, we will continue this professional life-changing conversation about propensity modeling to a whole new level.Â
We'll learn about the extensions and possibilities available to us to apply propensity modeling. (You are going to love our use case of using propensity models to identify how much to pay each Influencer, for your social influence campaigns!)Â
I'll share the challenges Data Scientists need to be aware of, and how to solve them. And, elements to be careful about for interactions that will invariably happen between your propensity scores and the ad platform.
See you next Thursday.
| Bottom line.
While there are AI stories that seem fantastical or complete hype, propensity modeling is an example of how it is already giving Marketers superpowers. We can simplify all that it takes to acquire a high-value customer (or prevent churn!).
As we remove complexity from our lives, I'm excited that I get more time and imagination to spend on things that actually matter massively: Creative.
Carpe diem!
Avinash. | PS: Book Recommendation.If you are an entrepreneur, a wannabe entrepreneur (like me), or bring an entrepreneur's mindset to work every day (like me!)... My dear friend, and fellow co-founder of Market Motive, Michael Stebbins, has written a wonderful new book: Backward Entrepreneur.From defining and validating the most sustainable market, playbooks to think smart, move fast, acquiring your first 1,000 customers, and keys to earning the biggest payouts on exit... Michael makes the complex accessible. Check the book out here. | Committed to investing in your professional growth? Upgrade to TMAI Premium here - it is published 50x / year. | |
|
|
[ Subscribe] Â |Â Â [ Web Version]Â Â |Â [ Unsubscribe]
                            ©2022 ZQ Insights  |  PO Box 10193, San Jose, CA, 95157, United States of America
|
|
|
|